3. Signal conditioning and acquisition chains

An automotive impact sensor — whether a MEMS accelerometer in the airbag ECU, a satellite crash sensor on the B-pillar, or a pressure-based side-impact sensor in the door — produces an analog voltage that must be digitized before any calibration coefficient can be applied. The chain between the sensor element and the calibration algorithm is not transparent. Every stage — anti-aliasing filter, sample-and-hold, ADC, and the digital decimation that follows — shapes the signal and sets a hard ceiling on how accurately a sensitivity or offset coefficient can be determined. This section examines each stage and shows how its parameters propagate into calibration uncertainty.

Anti-Aliasing Filter: The First Line of Defense

Impact signals contain energy well above the band of interest. A frontal crash pulse has meaningful content from DC to roughly 300–500 Hz, but the sensor's mechanical resonance and the electrical noise from the vehicle harness can extend into the tens of kilohertz. If the ADC samples at fs, any energy above fs/2 folds back into the baseband and becomes indistinguishable from the crash signal. An anti-aliasing filter must attenuate these components before the sample-and-hold.

In practice, a second-order or fourth-order Butterworth low-pass filter is common. For a target signal band of 500 Hz and a sample rate of 2 kSPS, the Nyquist frequency is 1 kHz. A second-order Butterworth at 500 Hz provides only about 12 dB attenuation at 1 kHz — insufficient if out-of-band energy is strong. A fourth-order filter at the same corner frequency yields roughly 24 dB at 1 kHz, which is usually acceptable for airbag-grade accelerometers. The filter corner frequency must be chosen so that passband ripple and group-delay variation do not distort the crash pulse shape used for calibration.

Sample Rate and Decimation

The sample rate determines both the Nyquist limit and the noise bandwidth. A higher fs spreads the same broadband noise power over a wider spectrum, reducing the in-band noise density after decimation. However, higher fs also increases data throughput and power consumption, which matters for satellite sensors on a LIN or PSI5 bus.

Consider a sensor with a noise density of 100 µg/√Hz. If the signal band is 500 Hz, the integrated RMS noise is 100 µg/√Hz × √500 Hz ≈ 2.24 mg RMS. If the sample rate is 2 kSPS and a digital decimation filter reduces the effective bandwidth to 500 Hz, the noise floor at the calibration output is set by that 500 Hz band. Doubling fs to 4 kSPS without changing the decimation ratio does not reduce in-band noise; it only relaxes the analog anti-aliasing filter requirement. The calibration accuracy depends on the effective noise bandwidth, not the raw sample rate.

Oversampling followed by decimation is a common technique. If the ADC runs at 16 kSPS and the decimation filter outputs 2 kSPS, the oversampling ratio (OSR) is 8. Each doubling of OSR improves the SNR by approximately 3 dB for a first-order modulator, or 6 dB for a second-order sigma-delta. This directly reduces the noise floor that limits offset and sensitivity calibration.

ADC Resolution and Quantization Noise

An N-bit ADC has a quantization step of LSB = Vref / 2^N. The quantization noise power, assuming a uniform distribution, is q²/12, where q is the LSB size. For a 12-bit ADC with a 5 V reference, LSB = 1.22 mV. The RMS quantization noise is approximately 0.35 mV. If the sensor sensitivity is 20 mV/g, this corresponds to 0.0175 g RMS — a significant fraction of a typical offset calibration tolerance of ±0.05 g.

Increasing to 16 bits reduces the LSB to 76 µV and the quantization noise to 22 µV RMS, or 0.0011 g RMS. The quantization noise is then negligible compared to the sensor's own noise floor. In practice, the ADC resolution should be chosen so that quantization noise is at least 10 dB below the sensor noise floor; otherwise, the ADC becomes the limiting factor in calibration repeatability.

ADC Resolution	LSB (Vref = 5 V)	Quantization Noise (RMS)	Equivalent Noise (g, 20 mV/g)
10-bit	4.88 mV	1.41 mV	0.070 g
12-bit	1.22 mV	0.35 mV	0.0175 g
14-bit	305 µV	88 µV	0.0044 g
16-bit	76 µV	22 µV	0.0011 g

Note: Values are calculated from ideal quantization noise models and assume a uniform input distribution. Actual performance depends on ADC architecture, reference stability, and input-referred noise.

Noise Floor and Calibration Accuracy

The total noise floor at the calibration input is the root-sum-square of sensor noise, analog front-end noise, and quantization noise. For a typical airbag accelerometer with 100 µg/√Hz noise density and a 500 Hz bandwidth, the sensor noise alone is 2.24 mg RMS. If the calibration procedure averages M samples, the noise on the averaged result is reduced by √M. Averaging 100 samples reduces the noise to 0.224 mg RMS, which is well within a ±0.05 g offset tolerance.

However, averaging only helps if the noise is white and the signal is stationary during the averaging window. In a vehicle crash test, the signal is non-stationary; calibration is typically performed on a shaker table with a sinusoidal or step input, where averaging over multiple periods is valid. The calibration engineer must verify that the noise floor is low enough that the standard error of the estimated sensitivity coefficient is below the required tolerance.

Signal Chain Visualization

The following diagram shows the complete signal conditioning and acquisition chain for a typical impact sensor calibration setup, from the sensor element through the anti-aliasing filter, sample-and-hold, ADC, and digital decimation, with the noise contributions at each stage.

Practical Calibration Example

Suppose a satellite impact sensor is calibrated on a shaker table at 100 Hz with a 1 g peak sinusoidal input. The sensor sensitivity is nominally 20 mV/g. The signal chain uses a 16-bit ADC at 2 kSPS with a 500 Hz anti-aliasing filter. The calibration procedure acquires 200 cycles, averages the peak-to-peak amplitude, and computes sensitivity as S = Vpp / (2 × 1 g).

The noise floor at the ADC input is 2.38 mg RMS. After averaging 200 cycles, the noise on the amplitude estimate is 2.38 mg / √200 ≈ 0.168 mg. The sensitivity error due to noise is 0.168 mg / 1 g = 0.017%, which is negligible. The dominant error source is the shaker table's own harmonic distortion and the reference accelerometer's calibration uncertainty, not the signal chain.

If the same sensor were calibrated with a 10-bit ADC, the quantization noise would be 0.070 g RMS. After 200-cycle averaging, the noise on the amplitude estimate would be 0.070 g / √200 ≈ 4.95 mg, or 0.495% sensitivity error. This would exceed a typical 0.5% sensitivity tolerance, making the 10-bit ADC the limiting factor. This example illustrates why ADC resolution and noise floor must be evaluated together with the averaging strategy, not in isolation.

Summary of Design Parameters

To achieve calibration accuracy targets, the signal conditioning chain must satisfy three conditions: (1) the anti-aliasing filter must attenuate out-of-band energy below the quantization noise floor at fs/2; (2) the ADC resolution must be high enough that quantization noise is at least 10 dB below the sensor noise floor; and (3) the effective noise bandwidth after decimation must be matched to the calibration signal bandwidth, with averaging used to reduce the noise on the final coefficient estimate. When these conditions are met, the calibration accuracy is limited by the sensor and the reference instrumentation, not by the acquisition chain.